Real-time transient stability assessment using stacked auto-encoders

Abstract

Purpose This paper aims to report how one can assess the transient stability of a power system by using stacked auto-encoders. Design/methodology/approach The proposed algorithm works in a power system equipped with the wide area measurement system. To be more exact, it needs pre- and post-disturbance values of frequency sent from phasor measurement units. Findings The authors have investigated the performance of the proposed method. Going through details, the authors have simulated many contingencies, and then have predicted the transient stability in each of which by using the proposed algorithm. Originality/value The results demonstrate that the algorithm is fast, and it has acceptable performance under different circumstances including the change of system topology and failures of telecommunication channels.

Artificial Intelligence Techniques for Power System Transient Stability Assessment

“…Datasets are built around the tacit premise that the TSA can be represented, essentially, as a binary classification problem, where the distinction between stable and unstable cases needs to be clearly established. This is done by introducing a so-called transient stability index (TSI) [8]. Since the loss of stability in a power system is a low-probability event, the ensuing datasets will be class imbalanced, which will have important repercussions on the supervised training of machine learning models.…”

“…Hang et al employed principal component analysis [28]. Several papers proposed different kinds of autoencoders (stacked, denoising, variational), e.g., [8,16,29,30]. Mi et al in [22] proposed a special bootstrap method and the random selection of variables in the training process for tackling the curse of dimensionality.…”

Power System Transient Stability Assessment Using Stacked Autoencoder and Voting Ensemble

“…A fully automatic data processing pipeline follows (Section 2.2), with features engineering, statistical post-processing, stratified shuffle split and data scaling. Figure 2 presents a single-line diagram of the IEEE New England 39-bus test case power system that serves as a benchmark for the TSA analysis [6,29,30,37,42]. Power system features a total of 10 synchronous machines of different nominal powers, a number of transmission lines (TLs), three-phase power transformers and loads.…”

“…Hence, in order for the ML models to cope with this large amount of (mostly unlabeled) information it becomes necessary to reduce/filter the number of features, preferably using an unsupervised learning approach. This is seen as an indispensable data processing step of the classifier building pipeline [25,28,30,34]. Furthermore, since labeling of the contingency cases invariably produces (highly) imbalanced data sets, special care is needed during building and training of the subsequent ML models [36,37].…”

“…A hierarchical deep learning approach was presented in [28]. Autoencoders, as a special kind of ANN architecture, have also been applied to the TSA problem [29,30]. However, application of deep learning brings with it special problems of initialization, convergence, long training time, vanishing gradients, forgetfulness, dead neurons, and others [31,32].…”